upstream hicache fixes (#5570)
This commit is contained in:
@@ -571,6 +571,14 @@ class Req:
|
||||
self.prefix_indices, self.last_node = tree_cache.match_prefix(
|
||||
rid=self.rid, key=self.adjust_max_prefix_ids()
|
||||
)
|
||||
elif enable_hierarchical_cache:
|
||||
# in case last_node is evicted during scheduling, we need to update the prefix_indices
|
||||
while self.last_node.evicted:
|
||||
self.prefix_indices = self.prefix_indices[
|
||||
: -len(self.last_node.host_value)
|
||||
]
|
||||
self.last_node = self.last_node.parent
|
||||
|
||||
self.extend_input_len = len(self.fill_ids) - len(self.prefix_indices)
|
||||
|
||||
def adjust_max_prefix_ids(self):
|
||||
|
||||
@@ -489,6 +489,8 @@ class Scheduler(
|
||||
tp_cache_group=self.tp_cpu_group,
|
||||
page_size=self.page_size,
|
||||
hicache_ratio=server_args.hicache_ratio,
|
||||
hicache_size=server_args.hicache_size,
|
||||
hicache_write_policy=server_args.hicache_write_policy,
|
||||
)
|
||||
else:
|
||||
self.tree_cache = RadixCache(
|
||||
|
||||
Reference in New Issue
Block a user